Two categories
There is AI that demonstrates and AI that removes work. The first is easier to fund and easier to show. The second is the only one that appears in an operating metric.
The distinction is not about model quality. It is about whether the work was redesigned around the capability, or the capability was added to unchanged work.
Three tests before we ship an agent
It must be explainable to an auditor. It must be bounded to a named operational scope. It must be replaceable within a week. Agents that fail any of the three create a dependency the client cannot govern.
Those constraints are not caution. They are what makes an operations director willing to let software decide anything at all.
Measure hours, not accuracy
Accuracy is a precondition, not an outcome. The number that matters is how many hours of operational work have been removed, and what the team is doing with that time now. If nobody can answer the second half, the first half was theatre.